rsanek

Another case of study design not really supporting the headline. According to the paper, the comparison here was "no advice" or "AI advice": there was no "placebo" where you have advice / references from non-AI materials. Think e.g. Google's instant answer box, pre-AI. It's hard to say without the comparison, but I have to imagine displaying that would have similar effects.

There's plenty of other issues with the study, one of the primary ones being that they chose a relatively 'dumb' AI (Step 3.5 Flash), but also, specifically hand-selected wrong answers. People that are used to using competent AIs that are mostly correct would be operating off of experience that suggests they should trust AI. In this case, by design, they shouldn't have, but it's hard to fault the user for that.

What's been your experience using these models? In my experiments, while it is true that the models are less likely to outright refuse to answer "sensitive" questions, they are still very resistant to actually respond in a meaningful / useful way.

a Berkeley Haas study which describes how AI usage increases the intensity of work

The marketing around this 'study' has been impressive. They published the HBR article [0] and other promotional material [1] over 5 months ago, but as far as I can tell, the actual study itself (even a preprint version) still isn't available anywhere. Hard to evaluate how believable the results are without being able to actually see the details of how the study was done. For now, all we know is that it's based on ~40 qualitative interviews from a single company in a single industry.

[0] https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies...

[1] https://newsroom.haas.berkeley.edu/ai-promised-to-free-up-wo...

cost has nothing to do with it

customers will flock to the cheapest inference

If AA is to be believed then per-task it is about the same cost as Sol. Agree that it's very different from DeepSeek v4 Pro, which is ~15x cheaper than K3.

https://artificialanalysis.ai/models/kimi-k3#price-cost

AA results are out: https://artificialanalysis.ai/models/kimi-k3

Critically, that cost excludes the amortized amount the batteries will be over that period. And if you're cycling them every day like this, they're not going to last more than 15~20 years.

There’s a stew of factors at work behind the rise in consumer rage [including] the rise of AI customer service.

I'm not really sure the problem is AI here. I've had plenty of bad experiences with human customer service agents that were poorly-trained and hard to understand. An AI can be implemented in a shitty way, where it's at best a regurgitation of the help docs. Or, it can be implemented well, where it can autonomously take action without any human involved.

I have had a couple of experiences with the latter. Once you do, you realize that you actually don't need to talk to a person per se -- you just want your problem solved efficiently. That can be done with a human or an AI; it isn't really the most significant bit to be filtering on IMO.

I see tons of confustion in the comments on whether AI can or can't read it. Bit of a marketing miss -- they should have picked clearly different decoy vs. default actual messages.

Some people also use LLMs to generate flashcards. And of course, the result will be those impersonal, mediocre cards.

I won’t say LLMs are useless for this. But from my trials, I get about 1 card that’s useful to me out of 10, and even that 1 card still needs rewriting.

I don't know the specifics of how the author tried to do so, but from what I've seen the majority of attempts are, let me drop a chapter of a textbook and say "make flashcards." If that is what we are talking about, then yes, LLMs are useless.

In my mind, though, this is sort of like looking at the very first GitHub Copilot LLM autocomplete from a couple of years ago and concluding, yeah it's nice for one-liners, but it cannot write an app.

If you create a framework around your card-creation AI so that it can use tools, and verify its work to ensure common card-creation pitfalls don't happen, you can get pretty high-quality cards. In my experience, you go from a 10~20% hit rate to a ~90% hit rate, which in my mind is good enough. I got to ~75% quality just from a two extra LLM calls that would assess a potential card against a standard set of rules (adapted from [0]). There are huge Pareto gains to be had here.

I've generated thousdands of cards over the last few months this way. I let the AI add it directly to Anki via AnkiConnect. Then, if when I go to review I find a card that my AI created and I don't like it, I just delete it.

Removing the limitation of card creation is really quite compelling, and I think the area is still highly under-invested in. Would be cool to see a generic framework evolve that one could use. For now, I've been using a personal fork of clanki [1].

[0] https://supermemo.guru/wiki/20_rules_of_knowledge_formulatio...

[1] https://github.com/jasperket/clanki

It's a balance. Maybe a helpful analogy would be a book -- yes, reading a book is effortful, and yes, "almost everyone" does not read. Still, I think most wouldn't consider it crazy to say that reading can be "fun."

Based on the Intelligence vs. Cost graph, not clear to me why anyone would use Terra? Luna looks quite interesting though, happy to see OpenAI still serving the more budget-oriented side of the market (seems like Anthropic and Google have lost interest there).

https://artificialanalysis.ai/articles/gpt-5-6-has-landed

This doesn't seem that big of a deal to me? I mean, in any other area where I want an assessment of a product, I'm not going to trust what the product producer says about it at face value -- obviously they're going to be biased. This is the whole raison d'etre for independent testing, like https://artificialanalysis.ai.

So interesting to read about AI perceptions from the other side. In my mind, the problems with these drafts were moreso that the writers using AI could not (or would not) actually engage and improve them when given feedback, not necessarily that their writing process did not match the traditional one.

I'm not sure I see any inherent problem with publishing books written with the help of AI. As with software, I don't really care much how it's made, I care bout my experience with the finished product.

"Is it worth the $?" is ultimately the question that will be asked of anything one pays for, regardless of how exactly it was produced.

Interesting perspective. In my experience the risk is actually that it results in alert fatigue, which means that drivers that would otherwise pay attention to such an alert no longer do.

I rarely make cards by hand anymore. I would recommend forking something like https://github.com/jasperket/clanki and editing it (perhaps with an agent) so that it works exactly to your liking.

I think you're measuring one method against a goal it wasn't set for. I assume you are referring to Mammoth direct air capture in your comment -- Stratos in Texas will soon be up, at ~15x the size of Mammoth. But DAC is only one method, and it's the hardest to scale. Just look at options outside of DAC, like Vaulted Deep, whose costs (financially and energy-wise) are far lower.

I do agree that it's unlikely that we can ignore reduction and just depend on purely scaling capture, especially if we care about avoiding more negative climate effects as the scaling goes on. But to say it is "completely infeasible" is not accurate.

The exepense is what's not to like. Far cheaper to build a single-story warehouse + outside parking lot than a second story above a parking garage.

It's fair to call out issues with the tool. But I think for individuals searching for jobs, using LLMs as the scapegoat for why it's hard to find a role is not terribly helpful.

In my experience, cold-applying has always worked essentially as a black hole, and LLMs haven't changed that much. The reality is that alternative avenues are always necessary to get the job you want. That could be a third-party recruiter; reaching out to a hiring manager on LinkedIn; or using your network to get referrals. Those continue to work whether the company is using a bone-headed tool like this or not.

Let's play this out further. How about high school, should there be grades there? Tests at all levels also typically involve a grade / metric -- are those included too?

Do you mean very well? Lots of proof out there that it works; Fin, Sierra and others already operate on a value based pricing model where they only get paid if the AI actually resolves issues.

If you can't concentrate while people are working on computers near you, I don't think you'll do well in any workplace that is based around knowledge work.

The CCP's approach has historically been to subsidize their companies far more than other countries do. Why would LLMs be any different?

https://www.oecd.org/en/data/dashboards/magic-database-indus...

Turns out, BLS actually lists this stuff when they release CPI figures.

Used cars & trucks; butter; cheese; flour; chicken; textbooks; drugs are all down since ~2 years ago. Not an exhaustive list!

https://www.bls.gov/news.release/cpi.htm

That's just one part of it. According to https://openrouter.ai/docs/guides/routing/model-variants/exa...

  We use three classes of signals:
   * Tool-calling success and reliability from real traffic
   * Provider performance metrics such as throughput and latency
   * Benchmark and evaluation data as it becomes available

Not quite -- those were loans that were largely repaid, quite different from the subsidies CCP uses for its industrial base.

ProPublica has a nice tracker that shows the loans: https://projects.propublica.org/bailout/

You can also see how subsidies compare between China and OECD in this recent doc. In autos, China subsidizes to the tune of 2-3% of revenue vs. <0.5% for North America. https://www.oecd.org/en/data/dashboards/magic-database-indus...

Who's paying the $50k? I don't see how it makes sense to pay that much for a home-grown setup when I could pay <$5k/year total for both of the two best frontier models at effectively unlimited usage.

The number on the box gets bigger every release.

Not really tho right? Since we got to 1m context in mid 2025 nearly no one has gone higher.

While it's a little unstable, I've found Docker's sbx to be a great sandbox to run agents with --dangerously-skip-permissions

Perhaps an idea that could work is that if you're a lab that is releasing closed source models, you have to also release open source ones. gpt-oss is now old but was decent when it came out. Nemotron is solid, especially the recent ultra release. And Nvidia especially has a much better story vs Chinese models around releasing all parts (including pre and post training data), not just the model itself.